Method and device for making available a virtual sensor with machine learning

Abstract

The invention relates to a method for making available a virtual sensor designed to determine, according to an operating variable curve (u) of a technical system (200, 300), said operating variable curve having been determined in particular according to a measurement performed on a technical system (200, 300), a state variable (x) characterizing a state of the technical system, wherein: first training data (D 1) is made available which comprises pairs of time curves of the operating variable (u(t)) and associated curves of the state variable (x(t)), said pairs having been determined empirically (i.e. in particular according to measurements); according to said data, a first model (h 1) is determined which characterizes an evolution of the state variable (x(t)) according to the operating variable (u(t)) (and according to the state variable (x(t))); by means of this first machine learning system, second training data (D 2) is generated which comprises time curves of the operating variable (u(t)) and associated curves of the state variable (x(t)), said curves of the state variable having been determined by means of the first model (h 1); according to said second training data (D 2), a second model (h 2) is trained which characterizes an evolution of the state variable (x(t)) according to the operating variable (u(t)) (and according to the state variable (x(t))) and which is less complex than the first model (h 1); the virtual sensor is made available such that it determines the operating variable (x) according to the state variable (u) by means of the second model (h 2).

Description

[0001]R. 408505 Description Title Method and device for providing a virtual sensor with machine learning The invention relates to a method for providing a virtual sensor, the virtual sensor, a method for operating the virtual sensor, a computer program, a machine-readable storage medium, and a training device. State of the art From DE 102022200022 A1, a method for providing an aging state model for determining a modeled aging state SOH of a battery cell is known, comprising the following steps: ^ Providing training data sets, each of which assigns an empirically determined aging state to an operating variable profile of one or more operating variables at a specific point in time; ^ Providing a data-based aging model that is designed with a neural differential equation system,to model internal electrochemical states of the energy storage device, wherein the internal states are mappable to an aging state, wherein at least one of the differential equations of the differential equation system comprises a sum of a deterministic model term and a data-based correction term formed by a data-based correction model, ^ training the data-based aging model based on the training data sets so that model parameters of the data-based correction model R. 408505 are determined. It is then also known to use this aging state model as a virtual sensor. Therefore, a method for determining the modeled aging state (SOH) of the battery cell at a specific point in time is also known from the cited document.with the following steps: ^ Providing a business variable profile for one or more business variables; ^ Providing a trained, data-based aging state model based on a neural differential equation system and trained to associate a business variable profile with an aging state; ^ Determining or predicting the modeled aging state by evaluating the aging state model with the business variable profile in time steps. Neural differential equation systems are derived from Chen et al. Neural Ordinary Differential Equations,https: / / arxiv.org / pdf / 1806.07366.pdf known. Advantages of the Invention Virtual sensors generally refer to the determination of a profile of a variable characterizing a state of a technical system as a function of an operating variable profile of the technical system (determined as a function of a measurement). The invention with the features of independent claim 1 has the advantage over that known from the prior art that the virtual sensor is particularly powerful. In particular, it is robustly extrapolatable. Further aspects of the invention are the subject of the independent claims. Advantageous developments are the subject of the dependent claims. Disclosure of the Invention R. 408505 In a first aspect, the invention relates to methods for providing a virtual sensor which is configured to determine a variable as a function of a variable determined, in particular, as a function of a measurement on a technical system.Operating variable profile of a technical system to determine a state variable characterizing a state of the technical system, wherein first training data are provided which comprise empirically (i.e. in particular dependent on measurements) determined pairs of temporal profiles of the operating variable and associated profiles of the state variable, wherein depending on this data a first model is determined which characterizes an evolution of the state variable depending on the operating variable and the state variable, wherein by means of this first machine learning system second training data are generated which comprise temporal profiles of the operating variable and associated profiles of the state variable determined by means of the first model, wherein depending on this second training data a second model is trained which characterizes an evolution of the state variable depending on the operating variable (and the state variable), and which has a lower complexity,i.e., a smaller number of parameters that can be adjusted during training than the first model, whereby the virtual sensor is provided in such a way that it determines the operating variable depending on the state variable using the second model. This has the advantage that less real training data is required, while at the same time the extrapolation capability is improved. Furthermore, the reduction in complexity leads to a more resource-efficient virtual sensor. In a further development, it can be provided that the first model and / or the second model comprises a physical model and a machine learning system such as an artificial neural network or a Gaussian process. The physical model has the advantage of improved explainability, which is particularly important for safety-critical applications of the virtual sensor. In this modeling approach, the machine learning system fulfills the task ofcompensate for any modeling inaccuracies of the physical model to R. 408505. However, this data-based portion of the modeling is sometimes not as easy to interpret or explain. Therefore, in a further development, it can be provided that the first model and the second model comprise the same physical model. This means that the machine learning system of the second model is reduced in complexity compared to the machine learning system of the first model. This results in better explainability of the virtual sensor, since data-based terms of the machine learning system are converted into structured terms of the physical model. Furthermore, this results in a more robust prediction of the virtual sensor, thus an improved extrapolation capability. In a further development of this aspect, it can be provided that the second model comprises only the physical model,The machine learning system is thus reduced to a "model" without free parameters. Alternatively or additionally, it can be provided that the physical model of the second model is further reduced in complexity using model order reduction methods in order to achieve a better ratio of accuracy to model complexity. Standard model order reduction methods, such as Krylov subspace methods, can be used for this purpose. Advantageously, in these aspects, the first model and / or the second model describe the evolution of the state variable as a neural differential equation. Such neural differential equations are known, for example, from Chen et al., Neural Ordinary Differential Equations, 2019, https: / / arxiv.org / pdf / 1806.07366.pdf. These are particularly well suited for modeling dynamic systems and thus for virtual sensors.because they exhibit high model accuracy and robustness. The state variable can then advantageously be determined by integrating the neural differential equation. R. 408505 In further developments, it can be provided that the second model is selected from a plurality of candidate functions of varying complexity, i.e., with a different number of adjustable parameters, whereby the second model is selected from the plurality of candidate functions in a Pareto-optimal manner with regard to the model complexity of the candidate functions and the performance of the candidate functions. This makes it particularly easy to resolve the conflict of objectives between model quality and the number of parameters. In an exemplary concrete application, the technical system is an electrical energy storage device, in particular a battery.and the state variable characterizes in particular an aging state of the electrical energy storage device. In a further exemplary application, the technical system is an electrical machine and the state variable characterizes in particular a temperature of a rotor of the electrical machine. In a further aspect, the invention relates to a virtual sensor which is configured to determine the state variable characterizing the state of the technical system as a function of an operating variable profile of the technical system, in particular as a function of a measurement on a technical system, which was provided using one of the aforementioned methods. In a further aspect, the invention relates to a method for operating this virtual sensor, in which the operating variable profile of the technical system is determined, in particular measured,and the state variable characterizing the state of the technical system is determined. Embodiments of the invention are explained in more detail below with reference to the accompanying drawings. In the drawings: R. 408505 Figure 1 schematically shows an exemplary first application of a virtual sensor; Figure 2 schematically shows an exemplary second application of a virtual sensor; Figure 3 shows a flowchart of an embodiment of the method for providing a virtual sensor, with subsequent optional operation of the virtual sensor; Figure 4 schematically shows a device for carrying out the method for providing the virtual sensor. Figure 1 shows a motor vehicle (100) with a sensor (110), which in the exemplary embodiment is connected to an electrical energy storage device (200), in particular a battery, which can comprise at least one battery cell, or a fuel cell,measures the applied voltage. This electrical energy storage device (200) can, for example, be used to supply a drive unit (130) of the motor vehicle (100) with energy. A measured value from the sensor (110) is transmitted to a computer (120), which, using the virtual sensor provided with the method illustrated in Figure 3, determines a variable therefrom, which in some embodiments determines a state of health (SOH) of the electrical energy storage device (200). Depending on this determined state of health, it is possible for the drive unit (130) to be controlled accordingly. In some embodiments, it can be provided that the computer (120) is located in the motor vehicle (100). In alternative embodiments, it can be provided that the computer (120) is located outside the motor vehicle (100). For example, it is possiblethat the drive unit (130) is controlled with an optimal operating strategy. R. 408505 Alternatively or additionally, it is possible that, depending on the predicted aging state, maintenance measures on the electrical energy storage device (200) are predictively displayed. In other embodiments, in which the electrical energy storage device (200) is provided by a battery, it is possible that a charging strategy for the battery is optimally controlled. Figure 2 illustrates further embodiments for the application of virtual sensors. A motor vehicle (100) with a sensor (111) which, in the exemplary embodiment, measures a voltage applied to an electrical machine (300), in particular a generator or an electric motor, in particular a voltage,which is applied to a winding of a stator or to a winding of a rotor (310) of the electric machine (300). This electric machine (300) can, for example, be part of the drive unit (130) shown in Figure 1 and can receive energy from the electrical energy storage device (200) also shown in Figure 1. A measured value from the sensor (111) is transmitted to the computer (120), which, using the virtual sensor provided by the method illustrated in Figure 3, determines therefrom a variable that characterizes a temperature in a rotor (310) of the electric machine (300). Depending on this determined temperature, it is possible for an actuator (135) of the motor vehicle (100) to be controlled accordingly. In some embodiments, it can be provided that the computer (120) is located in the motor vehicle (100). In alternative embodiments, it can be providedthat the computer (120) is located outside the motor vehicle (100). In some embodiments, it is possible for a cooling strategy of the electric machine (300) to be adapted depending on the determined temperature. Figure 3 describes a flowchart of embodiments for providing and subsequently operating virtual sensors. R. 408505 First, the virtual sensor is provided. For this purpose, initial training data, provided (1000). These comprise a temporal profile of an operating variable (u) of a technical system, and an associated state variable (x) characterizing a state of the technical system. In the embodiments illustrated in Figure 1 and Figure 2, the operating variable (u) corresponds to the variable determined by the sensor (110, 111). These associated values ​​from profiles of the operating variable (u) and state variable (x) can, for example, have been determined on test benches. Then, a first model ℎ ^ ( ^ ( ^ ) , ^ ( ^ ) , ^, ^, ^ ) = ^ ( ^ ( ^ ) , ^ ( ^ ) , ^, ^ ) + ^ ( ^ ( ^ ) , ^ ( ^ ) , ^, ^ )provided (1100). This first model (ℎ ^ ) comprises, in some embodiments, a physical model (f), e.g., a 1- or 2-node model (e.g., rotor temperature model) or an electrochemical battery cell model. This physical model (f) determines an approximation of the state vector x depending on the operating variable (u) (for the battery illustrated in Figure 1, this can, for example, be curves of the operating variable (u) during cycles using requested currents for charging or discharging the battery). The physical model is parameterized with parameters ^. The first model (ℎ ^) further comprises a machine learning system (g), for example an artificial neural network or a Gaussian process. The machine learning system (g) describes the data-based part of the modeling, which R. 408505 serves to improve the purely physical modeling. The machine learning system is parameterized with parameters ϕ. This first model (ℎ ^ ) is calculated using the first training data ^ ^ now in some embodiments as a neural differential equation system trained in a known manner, ie according to the equation In some embodiments, the state variable (x) describes a state vector and includes physically measurable or interpretable variables such as the temperature of the rotor (310) of the electric machine (300) or characterizes a thickness of an SEI or a quantity of cyclable lithium in the battery (200). Artificial data is then generated using the thus trained first model (ℎ^ ) is generated (1200). For this purpose, curves of the operating variable (u) are provided, for example, based on load profiles, which in some embodiments were derived from real data (e.g., from real-life or test runs). Using the first model (ℎ ^ ), a corresponding curve of the state variable (x) is then determined and provided from these respective curves of the operating variable (u). This can be interpreted as an artificial label. The pairs of curves of operating variable (u) and state variable (x) thus generated are used as second training data. provided. In some embodiments, the courses of the operating variable (u) cover a very broad spectrum, so that the first model (ℎ ^ ) is forced to extrapolate. A very large set of second training data (^ ^) can thus be generated by simulation. R. 408505 In some embodiments it can be provided that curves of the state variable (x) are generated by sampling the curves of the state variable (x) measured in step (1000) with a higher temporal resolution. In some embodiments it can be provided that such measured curves of the state variable (x) are interpolated in order to determine simulated curves of the state variable (x). In particular this interpolation can take place on an abstracted level and in some embodiments can comprise a number of rapid charging cycles and / or charging currents of the battery (200) and / or a driving performance and / or an operating strategy of the motor vehicle (100) or intervention limits (such as de-rating temperatures). This interpolation can take place on the basis of a single parameter or in a combination of at least two parameters.This artificially generated data can now also be used as second training data (^. ^ ) or the second training data (^ ^ ) are added. In some embodiments, it is provided that the second training data (^ ^ ) more data than the first training data (^ ^ ). This is facilitated by the fact that the first model (ℎ ^ ) determined curves of the state variable (x) are easier to generate than actually measured curves of the state variable (x) and the first model (ℎ ^ ) the physical model adapted to a domain ^ ( ^ ) , ^, ^ ) and can therefore exhibit better extrapolation properties for certain operating variables than would be the case for a pure machine learning system. A second model ℎ ^ ( ^ ( ^ ) , ^( ^ ) , ^, ^, ^ ) = ^^ ( ^ ( ^ ) , ^ ( ^ ) , ^, ^ ) + ^^ ( ^ ( ^ ) , ^ ( ^ ) , ^, ^ ) provided (1300). The second model (ℎ ^ ) includes, like the first model (ℎ ^ ) a physical term (^ ^ ) and a second machine learning system (^ ^ ), which in some embodiments may be given as an artificial neural network and R. 408505 in some embodiments as a Gaussian process. This physical term (^ ^ ) is in some embodiments the same term as the physical term (^) of the first model (ℎ ^ ) In some embodiments, it may be provided that the second machine learning system (^ ^) is constant zero, so ^ ^ ( ^ ( ^ ) , ^ ( ^ ) , ^, ^ ) = 0. This means that during the subsequent training of the second model (ℎ ^ ) which in the first model (ℎ ^ ) is transferred – as far as possible – into the physical model (f). This ensures the highest possible interpretability of the second model (ℎ ^ ) and a better extrapolation capability is ensured. In alternative embodiments, it can be provided that the second machine learning system (^ ^ ) is a neural network with a lower capacity than the given machine learning system (^). This may mean that the architecture of the second machine learning system (^ ^) has fewer parameters, for example fewer neurons per layer and / or fewer layers than the architecture of the machine learning system (^). In some embodiments, it is possible for nonlinear activation functions of the machine learning system (^) to be used in the second machine learning system (^ ^ ) are replaced by linear activation functions, which can also be characterized by fewer parameters. (In other words, in some embodiments, this may mean: ^ has a smaller dimensionality than ^). In alternative embodiments, it may be provided that the second machine learning system (^ ^ ) is a Gaussian process with lower capacity than the machine learning system (^) given as a Gaussian process. This may mean that the second machine learning system (^ ^) has fewer inducing points than the first machine learning system (^). During the subsequent training of the second model (ℎ ^ ) causes the second machine learning system (^ ^ ) has a lower model capacity than the machine learning system (^) of the first model (ℎ ^ ). This means that the R. 408505 physical part (^ ^ ) in the second model (ℎ ^ ) to extract as much information as possible from the machine learning system (^) of the first model (ℎ ^ ). In addition, the (now simplified) second machine learning system (^ ^ ) can also help to compensate for any modeling errors for greater accuracy. In some embodiments, the second machine learning system (^ ^) includes an interpretable data-based model, such as symbolic regression, a linear feature combination, an xNN known from https: / / arxiv.org / pdf / 1806.01933.pdf, or a (preferably linear) combination of basis functions. This ensures the subsequent training of the second model (ℎ ^ ) a high accuracy with good explainability and thus good validation. Then the second model (ℎ ^ ) using the second training data (^ ^ ) is trained (1400), in some embodiments analogously as described above, the first model (ℎ ^ ) as a neural differential equation system Now the virtual sensor will comprehensively cover the second model (ℎ ^) is provided (1500). The virtual sensor is configured to supply a curve of the operating variable (u) supplied to it to the second model and to determine the curve of the state variable (x) by integration (in some embodiments, numerical integration) of equation (2) and to provide it at its output. In some embodiments, it can be provided that the second model (ℎ ^ ) is further compressed before the provision of the virtual sensor in order to make the virtual sensor usable in the most resource-efficient way possible. This can be done, for example, by performing a Krylov subspace method with moment adjustment. R. 408505 In some embodiments, in this case and / or during the provision of the second model (ℎ ^ ) it should be provided that a majority of candidates for the second model (ℎ ^ ) or the compressed second model (ℎ ^) are determined, and then the one that is Pareto-optimal is selected for the provision of the virtual sensor, in particular with regard to resource expenditure and performance. This concludes the method for providing the virtual sensor. Subsequently, in some embodiments, it can be provided that the virtual sensor (1600) is now operated, for example as illustrated in the embodiments in Figures 1 and 2. It is possible that steps (1000) - (1500) are carried out repeatedly, for example at predeterminable time intervals. The first model (ℎ ^ ) is then initially equal to the already trained second model (ℎ ^ ) is set, and advantageously an unchanged version of the thus initialized first model (ℎ ^) for the Pareto-optimal selection is presented as an option. Figure 4 illustrates a training device (400) configured to carry out the method shown in Figure 3 (with the exception of the optional step 1600). In some embodiments, the method is embodied as a computer program stored on a computer-readable storage medium (420) and includes instructions that cause a computer (410) of the training device (400) to carry out the method when the computer (410) runs the computer program.

Claims

R. 408505 Claims 1. Method for providing a virtual sensor which is set up to determine a state variable (x) characterising a state of the technical system depending on an operating variable profile (u) of a technical system (200, 300) determined in particular depending on a measurement on a technical system (200, 300), wherein first training data (^ ^ ) are provided, which comprise empirically (i.e., in particular, dependent on measurements) determined pairs of temporal courses of the operating variable (u(t)) and associated courses of the state variable (x(t)), whereby, depending on these data, a first model (ℎ ^ ) is determined, which characterizes an evolution of the state variable (x(t)) depending on the operating variable (u(t)) (and the state variable (x(t))), whereby by means of this first machine learning system second training data (^ ^) are generated, the temporal courses of the company size (u(t)) and assigned by means of the first model (ℎ ^ ) determined courses of the state variable (x(t)), whereby depending on these second training data (^ ^ ) a second model (ℎ ^ ) is trained, which characterizes an evolution of the state variable (x(t)) depending on the operating variable (u(t)) (and the state variable (x(t))), and which has a lower complexity than the first model (ℎ ^ ), whereby the virtual sensor is provided in such a way that it can be used by means of the second model (ℎ ^ ) determines the operating variable (x) as a function of the state variable (u).

2. Method according to claim 1, wherein the first model (ℎ ^ ) and / or second model (ℎ ^ ) includes a physical model (^) and a machine learning system (^). R. 408505 3. The method according to claim 2, wherein the first model (ℎ ^ ) and the second model (ℎ^ ) comprise the same physical model (^).

4. Method according to one of the preceding claims, wherein the second model (ℎ ^ ) comprises only the physical model (^).

5. Method according to one of the preceding claims, wherein the first model (ℎ ^ ) and / or the second model (ℎ ^ ) describe the evolution of the state variable (x(t)) as a neural differential equation.

6. Method according to one of the preceding claims, wherein the second model (ℎ ^ ) is selected from a plurality of candidate functions of different complexity, i.e. with different numbers of adjustable parameters, whereby the second model (ℎ ^) is selected from the plurality of candidate functions in a Pareto-optimal manner with regard to model complexity of the candidate functions and performance of the candidate functions.

7. The method according to one of the preceding claims, wherein the technical system is an electrical energy storage device (200), in particular a battery, and the state variable (x(t)) characterizes in particular an aging state (SOH) of the electrical energy storage device (200).

8. The method according to one of claims 1 to 6, wherein the technical system is an electrical machine (300) and the state variable (x(t)) characterizes a temperature of a rotor (310) of the electrical machine (300). 9.Virtual sensor which is configured to determine the state variable (x) characterizing the state of the technical system (200, 300) depending on an operating variable profile (u) of the technical system determined in particular depending on a measurement on a technical system (200, 300), which was provided with the method according to one of claims 1 to 6. R. 408505 10. Method for operating the virtual sensor according to claim 9, in which the operating variable profile (u) of the technical system (200, 300) is determined, in particular measured, and the state variable (x) characterizing the state of the technical system (200, 300) is determined.

11. Computer program which is set up and contains instructions which cause a computer to carry out the method according to one of claims 1 to 8 or 10 when the computer (410) executes the computer program.

12. Machine-readable storage medium (420) on which the computer program according to claim 11 is stored.

13. Training device (400), comprising a computer (410), which is set up to carry out the method according to one of claims 1 to 8.

Images